FORTA: Byzantine-Resilient FL Aggregation via DFT-Guided Krum

Abstract

Secure federated learning enables collaborative model training across decentralized users while preserving data privacy. A key component is secure aggregation, which keeps individual updates hidden from both the server and users, while also defending against Byzantine users who corrupt the aggregation. To this end, Jinhyun So et al. recently developed a Byzantine-resilient secure aggregation scheme using a secret-sharing strategy over finite-field arithmetic. However, such an approach can suffer from numerical errors and overflows when applied to real-valued model updates, motivating the need for secure aggregation methods that operate directly over the real domain. We propose FORTA, a Byzantine-resilient secure aggregation framework that operates entirely in the real domain. FORTA leverages Discrete Fourier Transform (DFT) codes for privacy and employs Krum-based outlier detection for robustness. While DFT decoder is error-free under infinite precision, finite precision introduces numerical perturbations that can distort distance estimates and allow malicious updates to evade detection. To address this, FORTA refines Krum using feedback from DFT decoder, improving the selection of trustworthy updates. Theoretical analysis and experiments show that our modification of Krum offers improved robustness and more accurate aggregation than standard Krum.